These skills guide an agent that documents a technology stack (stacks), designs or audits the internal architecture (architecture), generates a project skeleton (scaffold), and defines the agent team that works on the project (orchestration). Security reports are especially useful when their instructions can cause an agent to exceed the authorized scope, expose sensitive data, widen another agent's permissions, or present an unsupported security conclusion as established fact.

Examples include:

• Prompt injection in project files or retrieved content that overrides the user's request or expands authorized access.
• Reads or writes outside the project boundary, including through symlinks.
• Secret values or private project content appearing in tool output, reports or external requests.
• Execution of project scripts, dependency installation or deployment during a documentation-only task, or any command outside an approved plan.
• Agent definitions written by orchestration that grant more than the approved design, bypass approvals, change host settings, or carry instructions copied from project files.
• Misleading verification claims that could affect a security decision, such as reporting an unchecked dependency as safe or configured tests as passed.

An alternative technology preference or ordinary documentation defect can be reported through the normal issue or contribution workflow.

Report privately through GitHub's private vulnerability reporting. If that form is unavailable, open a minimal public issue requesting private coordination without technical details. State only the general category and affected revision, and agree on a private channel before sharing sensitive details.

The project does not promise a response time. The supported versions are the latest release and the latest commit on main.

Do not put credentials, private code, customer data, exploitable details about a live system or actual personal configuration paths in a public issue. Do not test a suspected credential or reproduce an access violation against another person's environment.

Once an appropriate channel is agreed, a useful report contains:

• The affected revision and agent environment, where known.
• The request, relevant skill instruction and observed behavior.
• A minimal reproduction using an isolated fixture and synthetic data.
• The expected boundary, actual impact and any conditions needed to reproduce.
• A proposed correction or mitigation, if available.

Preserve enough detail to explain the issue while redacting sensitive values. For a symlink reproduction, use a nonexistent external target and record the attempted access rather than pointing at real credential files.

Stop the affected workflow and contain the exposure. Revoke or rotate exposed credentials through their owner, review unintended writes or external transmissions, and preserve a redacted account of what happened. Deleting a report does not by itself invalidate a leaked credential.

The skill's instructions and validation reduce known failure modes; they do not establish a security certification or an exhaustive review. Documented controls, successful local checks and observed agent behavior are different forms of evidence.